AI Skill Improvement Loop

What it does

Chooses between three paths depending on the situation: polish a finished creative piece and regenerate it, 'heal' a skill by tracing defects in its output back to the rule that allowed them, or teach a new skill through evaluation-driven iteration. Keeps a continuous improvement loop over the skills themselves.

When to use it

• Noticing a skill repeatedly produces low-quality output
• Wanting to polish one specific finished creative piece
• Documenting a new skill from real test cases

Steps

1. Identify whether the issue is in the final piece or the rule that produced it
2. To polish: regenerate the piece with targeted adjustments
3. To heal: trace the defect back to the specific rule that allowed it
4. Fix the rule and validate with a test case
5. To learn: define clear pass/fail evaluations
6. Iterate the skill until it consistently passes the evaluations

The skill

---
name: skill-improvement-loop
description: A single entry point to improve creative output and the skills that produce it, picking the right path for the situation.
---

# AI Skill Improvement Loop

## 1. What it does
Chooses between three paths depending on the situation: polish a finished creative piece and regenerate it, 'heal' a skill by tracing defects in its output back to the rule that allowed them, or teach a new skill through evaluation-driven iteration. Keeps a continuous improvement loop over the skills themselves.

## 2. When to use it
- Noticing a skill repeatedly produces low-quality output
- Wanting to polish one specific finished creative piece
- Documenting a new skill from real test cases

## 3. Steps
1. Identify whether the issue is in the final piece or the rule that produced it
2. To polish: regenerate the piece with targeted adjustments
3. To heal: trace the defect back to the specific rule that allowed it
4. Fix the rule and validate with a test case
5. To learn: define clear pass/fail evaluations
6. Iterate the skill until it consistently passes the evaluations

## 4. Safety rules
Always work in draft mode: nothing is sent, published, or paid for without explicit client approval. Never invent data, metrics, or results. No passwords, tokens, or API keys are requested or shown in chat.

## 5. Deliverable
A clear, reviewable draft of the output (text, code, report, or file), ready for the user to approve before it becomes final.

For Claude Code, save it as ~/.claude/skills/skill-improvement-loop/SKILL.md. In Cursor, ChatGPT or another assistant, paste it as project instructions.